chiark - git - mdw - catacomb/blob - math/karatsuba.h

   1 /* -*-c-*-
   2  *
   3  * Macros for Karatsuba functions
   4  *
   5  * (c) 2000 Straylight/Edgeware
   6  */
   7
   8 /*----- Licensing notice --------------------------------------------------*
   9  *
  10  * This file is part of Catacomb.
  11  *
  12  * Catacomb is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU Library General Public License as
  14  * published by the Free Software Foundation; either version 2 of the
  15  * License, or (at your option) any later version.
  16  *
  17  * Catacomb is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  * GNU Library General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Library General Public
  23  * License along with Catacomb; if not, write to the Free
  24  * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
  25  * MA 02111-1307, USA.
  26  */
  27
  28 #ifndef CATACOMB_KARATSUBA_H
  29 #define CATACOMB_KARATSUBA_H
  30
  31 #ifdef __cplusplus
  32   extern "C" {
  33 #endif
  34
  35 /*----- Header files ------------------------------------------------------*/
  36
  37 #ifndef CATACOMB_MPW_H
  38 #  include "mpw.h"
  39 #endif
  40
  41 /*----- Normal arithmetic macros ------------------------------------------*/
  42
  43 #define UADD(dv, av, avl) do {                                          \
  44   mpw *_dv = (dv);                                                      \
  45   const mpw *_av = (av), *_avl = (avl);                                 \
  46   mpw _c = 0;                                                           \
  47                                                                         \
  48   while (_av < _avl) {                                                  \
  49     mpw _a, _b;                                                         \
  50     mpd _x;                                                             \
  51     _a = *_av++;                                                        \
  52     _b = *_dv;                                                          \
  53     _x = (mpd)_a + (mpd)_b + _c;                                        \
  54     *_dv++ = MPW(_x);                                                   \
  55     _c = _x >> MPW_BITS;                                                \
  56   }                                                                     \
  57   while (_c) {                                                          \
  58     mpd _x = (mpd)*_dv + (mpd)_c;                                       \
  59     *_dv++ = MPW(_x);                                                   \
  60     _c = _x >> MPW_BITS;                                                \
  61   }                                                                     \
  62 } while (0)
  63
  64 #define UADD2(dv, dvl, av, avl, bv, bvl) do {                           \
  65   mpw *_dv = (dv), *_dvl = (dvl);                                       \
  66   const mpw *_av = (av), *_avl = (avl);                                 \
  67   const mpw *_bv = (bv), *_bvl = (bvl);                                 \
  68   mpw _c = 0;                                                           \
  69                                                                         \
  70   while (_av < _avl || _bv < _bvl) {                                    \
  71     mpw _a, _b;                                                         \
  72     mpd _x;                                                             \
  73     _a = (_av < _avl) ? *_av++ : 0;                                     \
  74     _b = (_bv < _bvl) ? *_bv++ : 0;                                     \
  75     _x = (mpd)_a + (mpd)_b + _c;                                        \
  76     *_dv++ = MPW(_x);                                                   \
  77     _c = _x >> MPW_BITS;                                                \
  78   }                                                                     \
  79   *_dv++ = _c;                                                          \
  80   while (_dv < _dvl)                                                    \
  81     *_dv++ = 0;                                                         \
  82 } while (0)
  83
  84 #define USUB(dv, av, avl) do {                                          \
  85   mpw *_dv = (dv);                                                      \
  86   const mpw *_av = (av), *_avl = (avl);                                 \
  87   mpw _c = 0;                                                           \
  88                                                                         \
  89   while (_av < _avl) {                                                  \
  90     mpw _a, _b;                                                         \
  91     mpd _x;                                                             \
  92     _a = *_av++;                                                        \
  93     _b = *_dv;                                                          \
  94     _x = (mpd)_b - (mpd)_a - _c;                                        \
  95     *_dv++ = MPW(_x);                                                   \
  96     if (_x >> MPW_BITS)                                                 \
  97       _c = 1;                                                           \
  98     else                                                                \
  99       _c = 0;                                                           \
 100   }                                                                     \
 101   while (_c) {                                                          \
 102     mpd _x = (mpd)*_dv - (mpd)_c;                                       \
 103     *_dv++ = MPW(_x);                                                   \
 104     if (_x >> MPW_BITS)                                                 \
 105       _c = 1;                                                           \
 106     else                                                                \
 107       _c = 0;                                                           \
 108   }                                                                     \
 109 } while (0)
 110
 111 /*----- Binary polynomial arithmetic macros -------------------------------*/
 112
 113 #define UXOR(dv, av, avl) do {                                          \
 114   mpw *_dv = (dv);                                                      \
 115   const mpw *_av = (av), *_avl = (avl);                                 \
 116                                                                         \
 117   while (_av < _avl)                                                    \
 118     *_dv++ ^= *_av++;                                                   \
 119 } while (0)
 120
 121 #define UXOR2(dv, dvl, av, avl, bv, bvl) do {                           \
 122   mpw *_dv = (dv), *_dvl = (dvl);                                       \
 123   const mpw *_av = (av), *_avl = (avl);                                 \
 124   const mpw *_bv = (bv), *_bvl = (bvl);                                 \
 125                                                                         \
 126   while (_av < _avl || _bv < _bvl) {                                    \
 127     mpw _a, _b;                                                         \
 128     _a = (_av < _avl) ? *_av++ : 0;                                     \
 129     _b = (_bv < _bvl) ? *_bv++ : 0;                                     \
 130     *_dv++ = _a ^ _b;                                                   \
 131   }                                                                     \
 132   while (_dv < _dvl)                                                    \
 133     *_dv++ = 0;                                                         \
 134 } while (0)
 135
 136 /*----- That's all, folks -------------------------------------------------*/
 137
 138 #ifdef __cplusplus
 139   }
 140 #endif
 141
 142 #endif